{"paper":{"title":"SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Gradient-based weight saliency enables effective unlearning of data, classes, or concepts in both image classifiers and generators while approaching exact retraining performance.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Chongyu Fan, Dennis Wei, Eric Wong, Jiancheng Liu, Sijia Liu, Yihua Zhang","submitted_at":"2023-10-19T06:17:17Z","abstract_excerpt":"With evolving data regulations, machine unlearning (MU) has become an important tool for fostering trust and safety in today's AI models. However, existing MU methods focusing on data and/or weight perspectives often suffer limitations in unlearning accuracy, stability, and cross-domain applicability. To address these challenges, we introduce the concept of 'weight saliency' for MU, drawing parallels with input saliency in model explanation. This innovation directs MU's attention toward specific model weights rather than the entire model, improving effectiveness and efficiency. The resultant m"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"To the best of our knowledge, SalUn is the first principled MU approach that can effectively erase the influence of forgetting data, classes, or concepts in both image classification and generation tasks.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That gradient-derived weight saliency reliably isolates the parameters responsible for the forgetting data without introducing large unintended side effects on retained knowledge, an assumption tested only through the reported empirical gaps to exact unlearning.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"SalUn uses gradient-based weight saliency to achieve effective machine unlearning of data, classes, or concepts in image classification and generation, narrowing the gap to exact retraining.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Gradient-based weight saliency enables effective unlearning of data, classes, or concepts in both image classifiers and generators while approaching exact retraining performance.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"5c965883341be7a0b6bb0095e7a6479c2b6c33afe5ff91def20967b70d5428fd"},"source":{"id":"2310.12508","kind":"arxiv","version":5},"verdict":{"id":"2e920ccc-c607-4539-af0f-b50cb532c824","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T17:52:24.110584Z","strongest_claim":"To the best of our knowledge, SalUn is the first principled MU approach that can effectively erase the influence of forgetting data, classes, or concepts in both image classification and generation tasks.","one_line_summary":"SalUn uses gradient-based weight saliency to achieve effective machine unlearning of data, classes, or concepts in image classification and generation, narrowing the gap to exact retraining.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That gradient-derived weight saliency reliably isolates the parameters responsible for the forgetting data without introducing large unintended side effects on retained knowledge, an assumption tested only through the reported empirical gaps to exact unlearning.","pith_extraction_headline":"Gradient-based weight saliency enables effective unlearning of data, classes, or concepts in both image classifiers and generators while approaching exact retraining performance."},"references":{"count":206,"sample":[{"doi":"","year":2018,"title":"Sanity checks for saliency maps","work_id":"69988c5b-fe68-4e95-85a8-cc41ba100920","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Gradient surgery for one-shot unlearning on generative model, 2023","work_id":"2a489613-80f2-45a6-b034-ee0e5080b7c3","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2019,"title":"Nudenet: Neural nets for nudity classification, detection and selective censoring, 2019","work_id":"eafa1051-a7d3-4670-9b40-49712361e685","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Membership inference attacks from first principles","work_id":"0ffe7056-ca0e-4fcf-81a6-073d0c49bd83","ref_index":6,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2018,"title":"Grad- CAM ++: Generalized gradient-based visual explanations for deep convolutional networks","work_id":"539fbab8-5471-4ff5-ac20-cfbd0a8a5e95","ref_index":7,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":206,"snapshot_sha256":"a51c80c0fb437dbb4b6bc8643e329061ccfb291e69f796e06d24c79d4a4eaed1","internal_anchors":17},"formal_canon":{"evidence_count":2,"snapshot_sha256":"331bb906d55fb4e0578a30a09c5f2a026a5e8997cb8ac340d144014f96ebe73d"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}